Griptape Run: Audio Transcription

Voice memos, interview recordings, audio you generated in another node - this is how you get their words into ComfyUI as plain text. Griptape Run: Audio Transcription transcribes an audio file and hands you back a STRING you can feed into any downstream text node, including another agent prompt.

It's a niche node, sure. But once you need it, it's the only one in this pack for the job. It rounds out the audio story here - the pack pairs it with the Text-to-Speech node on the other side, so you can literally go voice → text → agent → text → voice in one graph.

How it works

The node accepts audio either as a ComfyUI AUDIO tensor (from other audio nodes) or as a file path. An AUDIO input gets written to a temp FLAC file first, then loaded via Griptape's AudioLoader and passed to an AudioTranscriptionTask. The transcription engine is OpenAI's whisper-1 by default, using your OPENAI_API_KEY - which means this node needs a key, unlike the local-model parts of the pack.

You can also connect an AUDIO_TRANSCRIPTION_DRIVER to use a different backend, or a config agent whose audio-transcription driver gets used instead. If the driver ends up being a dummy (i.e., nothing real configured), the node returns a friendly apology message instead of crashing - it's the pack's way of telling you to configure a driver.

The inputs

• audio - an AUDIO tensor from another node, or
• audio_filepath - a path to a file on disk. Give it one or the other; both empty and you get "There is no audio file."
• driver - optional AUDIO_TRANSCRIPTION_DRIVER override.
• config - optional agent whose transcription driver gets used.

Output: OUTPUT, the transcribed text.

Installing

Same pack. ComfyUI Manager → search "Griptape" → Install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/griptape-ai/ComfyUI-Griptape

Restart ComfyUI. Deps: griptape[all], python-dotenv. You'll want an OPENAI_API_KEY in the sidebar Settings → Griptape (or .env) for the default whisper path.

Common gotchas

The torch thing, as ever: the pack pulls torch in as a dependency and Nvidia users sometimes need to reinstall it against the right CUDA (pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu121) after import errors. And the stale-griptape ImportError: cannot import name 'OllamaPromptDriver' is fixed with python -m pip install griptape -U.

Two transcription-specific notes: transcription cost scales with audio length on the API path (whisper bills per minute), and if you feed it a huge file you may want a real driver configured rather than the default. Also, unsupported audio formats will fail in AudioLoader before the task even runs - convert first if you get a load error.